Generative engine optimization (GEO) is the practice of making your pages and your brand more likely to be found, quoted and cited when an AI system such as ChatGPT, Perplexity, Gemini or Google's AI Overviews writes an answer. The name comes from a 2023 research paper. In practice it's SEO plus citation-focused writing and reputation work.
Generative engine optimization, defined
Generative engine optimization (GEO) is the work of getting an AI answer engine to pick your content as a source, and to name and describe your business correctly when it answers. It’s the core of our GEO and AI search optimization work, and it’s a narrower, more practical discipline than most of the marketing around it suggests.
A “generative engine” is any system that takes a question, finds sources, and writes the answer with a language model. ChatGPT with search switched on works this way. So do Perplexity, Microsoft Copilot, Gemini, and Google’s AI Overviews and AI Mode. Google describes its version as a “query fan-out” technique, issuing multiple related searches across subtopics and data sources and then writing from what comes back.
Being “in” one of those answers can mean four quite different things:
- Named: your business appears in the text, perhaps as a recommendation.
- Cited: a link to your page sits next to a claim, so the reader can click through.
- Used without credit: your page shaped the answer but someone else got the link, or nobody did.
- Described: the answer says something about you, and it may be out of date or wrong.
GEO tries to move you toward the first two and fix the fourth. It isn’t a file you install or a placement you buy. Nobody sells organic slots in ChatGPT’s answers.
Where the term came from: the GEO paper
The name comes from GEO: Generative Engine Optimization, by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. It was first posted in November 2023, revised twice in 2024, and presented at KDD 2024, a major data-mining conference. Most GEO articles quote one number from it and skip the method, which is a shame, because the method tells you how far to trust the number.
The authors built a benchmark called GEO-bench: 10,000 queries drawn from nine sources, including MS MARCO, Natural Questions, ELI5 and Perplexity.ai’s Discover feed, tagged by subject and query type. Their main test engine used GPT-3.5-turbo to answer each query from the top five Google results. They then rewrote one of those five sources using a given method and measured how much more of the answer came from it.
Visibility was scored two ways. One counted the words in the answer attributed to the source, weighted toward the start of the answer. The other was a model-judged score for how relevant, influential and prominent the source felt to a reader. The headline finding, in the abstract, is that their methods can boost visibility by up to 40% in generative engine responses.
The nine methods they tested
| Method | What it changes on the page | How it did |
|---|---|---|
| Quotation addition | Adds relevant quotes from named people or sources | Among the best |
| Statistics addition | Swaps vague claims for numbers | Among the best |
| Cite sources | Adds citations to credible references | Among the best |
| Fluency optimization | Smooths the writing | Smaller gains |
| Easy-to-understand | Simplifies the language | Smaller gains |
| Authoritative | Makes the tone more confident and persuasive | Smaller gains, varied by subject |
| Technical terms | Adds domain vocabulary | Smaller gains, varied by subject |
| Unique words | Adds distinctive vocabulary | Smaller gains |
| Keyword stuffing | Repeats query keywords | Worse than the unmodified page |
On the main benchmark, quotation addition scored 27.8 on the word-count measure against 19.5 for unmodified pages, and keyword stuffing scored 17.8, per the paper’s results. That last row is the one SEO people should sit with. The oldest trick in the book made pages less visible to a generative engine.
The authors repeated the experiment on Perplexity.ai with a smaller sample, and the pattern broadly held: quotations and statistics helped, keyword stuffing didn’t.
Lower-ranked pages gained the most
This is the part we find most interesting. Citing sources lifted visibility for the page ranked fifth in the search results by 115.1%, while the top-ranked page’s visibility fell by 30.3% on average, according to the GEO study.
Read that carefully. In a benchmark, a well-sourced page from further down the results took share from the page at the top. If even part of that holds in live engines, a smaller company that writes carefully has a real opening against a bigger one that doesn’t.
Results changed by subject
The paper is clear that no method wins everywhere. An authoritative tone helped most on debate and history questions. Statistics helped on law, government and opinion queries. Quotations worked for people-and-society and history topics, and citing sources helped on factual statements and law. If you sell to lawyers, numbers and references carry more weight than polished prose.
Where the study stops
We use this paper as writing guidance, not as a description of how ChatGPT or Google ranks sources. Three limits matter:
- It’s a 2023 benchmark. The test engine was GPT-3.5 reading five Google results. Today’s engines fan out into several searches, use their own crawlers and indexes, personalise answers, and run on much larger models.
- It assumes you were already retrieved. Every page in the experiment was in the top five. Most real GEO problems we see happen before that point: a blocked crawler, a page Bing never indexed, a business that isn’t mentioned anywhere the engine looks.
- It edits one page while the rest stand still. In a real market, competitors rewrite their pages too, and third-party sites (directories, reviews, press) don’t appear in the experiment at all.
So the finding to keep is the direction, not the size. Specific, sourced, quotable writing helps. Stuffing doesn’t. The percentages belong to the benchmark.
What works in practice
Here’s the order we’d work in, starting with whatever is broken.
Let the engines in
Each assistant fetches pages with its own crawler or through a search index. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. Perplexity says PerplexityBot surfaces and links websites in its search results and isn’t used to train foundation models. Check robots.txt, and check your CDN or firewall, which blocks AI bots by default on some setups without anyone noticing.
Get indexed in Google and Bing
Google’s AI features draw on Google’s index. Copilot draws on Bing’s. Verify both sites in their webmaster tools and submit sitemaps. Bing also supports IndexNow, a ping that tells participating engines (Bing, Naver, Seznam.cz, Yandex and Yep) when a page changes. Google doesn’t participate.
Write passages worth quoting
This is where the paper earns its keep. Compare two sentences from a hypothetical clinic page:
“We are a trusted provider of quality physiotherapy services for all your needs.”
“We treat sports injuries and post-surgery rehabilitation at our Dubai Marina clinic, with sessions in English and Arabic.”
The first gives an engine nothing to quote. The second answers three questions a patient might ask an assistant: what you treat, where, and in which languages. Add a named source or a real figure where you have one, and put the direct answer straight under a clear question.
Get mentioned where the engines already look
Run your buyers’ questions through the assistants and note which pages they cite. Those pages, often directories, comparison articles and trade press, become your outreach list. Our ChatGPT ranking walkthrough covers the prompt method step by step.
Make your business unambiguous
Same name, same services, same locations everywhere: your site, Organization structured data, Google Business Profile, LinkedIn and directories. For companies in the Gulf, match the Arabic and English names deliberately, because an engine merging sources can’t guess they’re the same firm.
Measure the right things
Google counts AI Overviews and AI Mode traffic inside the normal Search Console Performance report, “within the ‘Web’ search type”, so you can’t separate it there. Bing Webmaster Tools has an AI Performance report counting citations in Copilot and Bing’s AI summaries. For ChatGPT, Gemini and Perplexity, the reliable method is still a fixed prompt set, run repeatedly.
What’s hype
Some GEO claims don’t survive contact with the primary sources.
“We guarantee placement in ChatGPT.” Nobody can. Answers change between runs and between users, and there’s no organic slot for sale.
“You need an AI file or special schema.” Google says you don’t need to create “new machine readable files, AI text files, or markup” for its AI features, and that there’s no special schema.org markup to add. Structured data still helps engines understand a page, as long as it matches what’s visible. The llms.txt proposal is a separate question, and where llms.txt actually stands has its own guide.
“GEO replaces SEO.” The engines retrieve through search. We explain the overlap in GEO vs SEO vs AEO.
“Your AI visibility score is 37.4.” Tools that report a precise share of voice are sampling a handful of runs of a handful of prompts. The trend over months is useful. The decimal place isn’t.
Hidden instructions for AI. Text on a page telling models to recommend you is the hidden-keyword trick of twenty years ago in a new outfit. We wouldn’t risk a client’s domain on it.
What nobody knows yet
Plenty, and it’s better to say so.
None of the major engines publishes how it chooses among the pages it retrieves. Nobody outside those companies can say how much weight a brand mention without a link carries. OpenAI’s help page on ChatGPT search refers to third-party search providers alongside its own crawling, and Microsoft said in 2023 that Bing would be ChatGPT’s default search experience, but the current mix isn’t public. We also don’t know how long a citation lasts once you’ve earned it, or, for any given topic, whether an answer leans more on training data or on live search.
Our response to all of that is to measure rather than guess: the same prompts, the same engines, on a schedule, with the cited sources logged. For businesses here, that includes running prompts in Arabic as well as English, which is part of how we approach AI search optimization for Dubai clients.